[AMDGPU] Fix number of reserved SGPRs on CI to reflect flat scratch use

[rampitec]

Summary

We are using flat scratch on CI as well, so we need to reserve registers for that.

[• tstellarAMD]

lib/Target/AMDGPU/SIRegisterInfo.cpp

1125

We should be checking for SEA_ISLANDS here, since there is no flat_scratch on SI.

[rampitec]

Added explicit check for SEA_ISLANDS.

[arsenm]

lib/Target/AMDGPU/SIRegisterInfo.cpp

1126

This can probably be skipped if there are no stack objects

[rampitec]

lib/Target/AMDGPU/SIRegisterInfo.cpp

1126

Doesn't same apply to VI?

[arsenm]

I think D27150 already is trying to do this

[arsenm]

I think D27150 already is trying to do this

lib/Target/AMDGPU/SIRegisterInfo.cpp

1126

I think for VI because of the order, xnack requires reserving the 2 for flat_scratch even though it's unused. We can probably make flat_scratch allocatable in that case to avoid losing 2 registers

[rampitec]

In D27225#613907, @arsenm wrote:

I think D27150 already is trying to do this

It does, yes. I see however a test where 104 SGPRs are allocated on CI which exceeds the limit. MFI.hasFlatScratchInit() returns false, and whole spilling went to VGPRs, but we still mark scratch as used. I suppose that is already a separate problem of not releasing scratch.

[nhaustov]

Looks good as short term solution.

lib/Target/AMDGPU/SIRegisterInfo.cpp

 unsigned SIRegisterInfo::getNumReservedSGPRs(const SISubtarget &ST) const {
   if (ST.getGeneration() >= AMDGPUSubtarget::VOLCANIC_ISLANDS)
     return 6; // VCC, FLAT_SCRATCH, XNACK.
-  return 2; // VCC.
+  return 4; // VCC, FLAT_SCRATCH.
 }
 
 unsigned SIRegisterInfo::getMinNumSGPRs(const SISubtarget &ST,